Int. J. Mol. Sci. 2026, 27(10), 4349; https://doi.org/10.3390/ijms27104349 - 13 May 2026
Viewed by 367
Abstract
►
Show Figures
Artificial intelligence has advanced from merely predicting static protein structures to modeling equilibrium conformational ensembles. It now concurrently forecasts structure and binding affinity and actively participates in candidate selection during the initial stages of drug discovery. Foundation models introduced between 2024 and 2026,
[...] Read more.
Artificial intelligence has advanced from merely predicting static protein structures to modeling equilibrium conformational ensembles. It now concurrently forecasts structure and binding affinity and actively participates in candidate selection during the initial stages of drug discovery. Foundation models introduced between 2024 and 2026, including BioEmu, AlphaFlow, DiG, Boltz-2, Chai-1, NeuralPLexer, and explicit-solvent prediction systems such as SuperWater, have begun to address issues previously identified as fundamental concerns in earlier critiques of AI in drug discovery. Nevertheless, many of these models are presently accessible only as preprints and require validation through independent peer review. Evidence indicates a shift in the primary bottleneck from representation challenges to validation difficulties. However, this transition remains incomplete and heavily dependent on context. The risks associated with AI-enabled drug discovery are increasingly not solely about the models’ capacity to accurately represent ensembles, but also about whether the evidentiary standards used to validate AI-derived predictions keep pace with the rapidity with which these predictions are generated and employed. This article introduces a four-tier validation framework designed to align the extent of computational and experimental evidence with the translational and regulatory risks associated with various artificial intelligence (AI) applications within the molecular sciences. These applications include machine learning (ML) models that analyze sequences, structures, conformational ensembles, protein–ligand complexes, and molecular dynamics trajectories. Tier 1 addresses the internal reproducibility of ML inference when applied to molecular inputs; Tier 2 pertains to the robustness of molecular-science benchmarks such as CASP, CASF-2016, PoseBusters, and OpenFE; Tier 3 involves prospective experimental validation against biophysical and biochemical measurements; and Tier 4 encompasses clinical and translational calibration within physiologically based pharmacokinetic (PBPK) and quantitative systems pharmacology (QSP) frameworks. This validation hierarchy functions as an explicit conceptual guide, serving as a framework rather than a regulatory requirement. It is firmly grounded in established principles derived from ICH Q8/Q9/Q10, the FDA model-informed drug development (MIDD) approach, the EMA reflection paper on AI in the medicinal product lifecycle, and the EU AI Act. The manuscript further incorporates recent evidence from ensemble-aware AI, prospective docking, free-energy campaigns, and clinical-stage AI-derived candidates. It concludes with specific recommendations pertaining to lifecycle governance, uncertainty reporting, and the adoption of harmonized evidentiary templates for AI/ML applications in the molecular sciences.
Full article
Figure 1
